Liu Yao

@liuyao12

"blogs" at on interactive math, or just math

Vrijeme pridruživanja: prosinac 2010.

Tweetovi

Blokirali ste korisnika/cu @liuyao12

Jeste li sigurni da želite vidjeti te tweetove? Time nećete deblokirati korisnika/cu @liuyao12

  1. Prikvačeni tweet
    29. lis 2019.

    Some preliminary results (10-layer modified ConvNet, followed by a ReLU + softmax): not that I can see discernible rotation/scaling in the feature maps. Next, should try to suppress the "second-order derivatives" in the kernels.

    Prikaži ovu nit
    Poništi
  2. 28. sij

    TIL what “groups” (and “cardinality”) in Conv2d are, which would have simplified my code. Hate that they use well-established math terms when it simply means the matrix (whose entries are 3x3) is block diagonal. Block has come to mean something else in ConvNet architecture.

    Prikaži ovu nit
    Poništi
  3. 24. sij

    Note to self: Not following Senate impeachment trial, but was so moved by closing remark on Day2, that every1 knows we can’t count on Trump to choose country over personal gain. Thought: what if jury vote were anonymous? It’s the only way America can redeem itself.

    Prikaži ovu nit
    Poništi
  4. 7. sij

    The NYT report gave lots of actual details. Impulsive and defensive he was/is (classic Trump), but not for distracting impeachment it appears. Still, a total lack of considering the consequences which are being played out.

    Prikaži ovu nit
    Poništi
  5. 6. sij

    When will we see Republicans blame Pelosi and Schiff for destabilizing the Middle East? They pushed for the impeachment, and Trump in his natural reflex did the only thing he knew in re-election year.

    Prikaži ovu nit
    Poništi
  6. 3. pro 2019.

    To check if one layer is "close" to another (for regularization, say), we can't just direct-compare the 16×16 matrices, for they could be very close up to permuting the rows/columns. We'd need to compare all the coefficients of the characteristic polynomial, of which det is one.

    Prikaži ovu nit
    Poništi
  7. 3. pro 2019.

    So, to do Gaussian elimination (or take det) on that, we'd use convolution product (which thankfully is commutative). E.g., 3×3 convolve 3×3 is a 5×5 kernel. To get an idea what the Conv layer is doing, we may take the 16×16 determinant, & visualize the resulting large kernel 😱

    Prikaži ovu nit
    Poništi
  8. 3. pro 2019.

    As a classic example, the wave equation u_tt=Δu can be turned into a system of two equations (to turn back, we can do a sort of Gaussian elimination). A typical CNN may have 16 or more channels, so that'd be a 16×16 matrix of 3×3 kernels (differential operators)

    Prikaži ovu nit
    Poništi
  9. 26. stu 2019.

    Ok, a dirty trick to get only first-order operator is to symmetrize the 3×3 kernel (symmetric wrt the center). Now I see the multiple channels in CNN as a *system* of PDEs, or a matrix-valued differential operator (or in my mind, a differential-operator-valued matrix). Update👇

    Prikaži ovu nit
    Poništi
  10. 20. stu 2019.

    My prediction for the coming year: No idea what will come out of the impeachment process, or who is going to win in November. All I know is that it will get uglier (and we will miss the innocent days of 2016). Democracy is dead, again. This time there is no Sparta.

    Prikaži ovu nit
    Poništi
  11. 14. stu 2019.

    Any help with testing would be greatly appreciated :) (Still no suppression on "second-order" operators, and no "energy" regularization. This is the "minimalist" version, so it can be easily added to any existing ResNet.)

    Prikaži ovu nit
    Poništi
  12. 14. stu 2019.

    FWIW, here's implementation in PyTorch for DL newbies & researchers alike, based on the classic ResNet. It appears that ResNet may have already reached its capacity on CIFAR10, at 95%, but the same "twist" can be easily applied to any ResNet-based model.

    Prikaži ovu nit
    Poništi
  13. 8. stu 2019.
    Poništi
  14. 4. stu 2019.

    I put my Colab notebook out, and had a chance to pitch it. If to rewrite on Twitter it’d be several tweets long, so 👇 (Apologies for my frustration at self-promotion.)

    Prikaži ovu nit
    Poništi
  15. 26. lis 2019.

    Finally did some testing on MNIST... more works needed to make it work

    Prikaži ovu nit
    Poništi
  16. 23. lis 2019.

    It’s only fitting to do the same to Xi Jinping, but I can’t separate what is Xi asserting himself and what is people around him currying favors (as eunuchs did in late Han and late Ming). That’s the biggest mystery.

    Prikaži ovu nit
    Poništi
  17. 23. lis 2019.

    With all the leaks and what Trump spouts out daily, most of the craziness “made sense”. (Trump tower meeting? You bet.) Thus, my top three are: 1) How did Tillerson get his job as sec of state? 2) Jeffrey Epstein’s death. 3) Why did Trump withdraw from supporting the Kurds?

    Prikaži ovu nit
    Poništi
  18. 23. lis 2019.

    As fun as it is to ponder current US politics against the long history of imperial China (Trump = the idiotic son to ascend the throne), here’s a new one: what’s the Trump equivalent of the three 疑案 of late Ming (palace intrigues)?

    Prikaži ovu nit
    Poništi
  19. 16. lis 2019.

    Conspiracy theory of the day: Putin and/or Erdorgan are foreseeing the end of Trump and decide to cash in whatever they have on him, which they’d expose anyway to further plunge US into humiliation and constitutional chaos.

    Poništi
  20. 9. lis 2019.

    Just found that it was tweeted before, but no mention of Bai Juyi at the time.

    Prikaži ovu nit
    Poništi

Čini se da učitavanje traje već neko vrijeme.

Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.

    Možda bi vam se svidjelo i ovo:

    ·